On Mon, Nov 07, 2022 at 10:31:21AM +0000, Richard Biener wrote:
> On Mon, 7 Nov 2022, Jakub Jelinek wrote:
>
> > On Mon, Nov 07, 2022 at 10:02:11AM +0100, Richard Biener wrote:
> > > The following adds a fold_builtins pass at -O0, keying off some
> > > unwanted optimization and setting pointer alignment of the result
> > > of __builtin_assume_alignment before removing the call. This
> > > allows libatomic calls to be elided at -O0 on s390 for
> > >
> > > __uint128_t foo(__uint128_t *p)
> > > {
> > > return __atomic_load_n((__uint128_t *)__builtin_assume_aligned (p, 16),
> > > 0);
> > > }
> > >
> > > not sure how to reliably test this though.
> > >
> > > Thoughts?
> >
> > Do we really need a separate pass for it?
> > Can't we do it say during gimplification?
>
> gimplification would be too early for always inline - of course since
> we don't do any copy propagation the source pattern this works reliably
> are limited, mostly when used directly as arguments like in the example
> above.
Yeah, that was exactly my thinking, because we don't copy propagate at -O0,
it will only handle cases where there is exactly one SSA_NAME involved.
The advantage of doing it at gimplification time is that we don't need to
add an extra pass for -O0.
> So yes, the specific case in question would work when we elide
> __builtin_assume_aligned during gimplification at -O0 (or during
> the GIMPLE lower pass).
>
> Would you prefer that? Richard, would that work for you?
Jakub
